1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/dirtyCardQueue.hpp" 27 #include "gc/g1/g1BlockOffsetTable.inline.hpp" 28 #include "gc/g1/g1CardTable.inline.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1ConcurrentRefine.hpp" 31 #include "gc/g1/g1FromCardCache.hpp" 32 #include "gc/g1/g1GCPhaseTimes.hpp" 33 #include "gc/g1/g1HotCardCache.hpp" 34 #include "gc/g1/g1OopClosures.inline.hpp" 35 #include "gc/g1/g1RootClosures.hpp" 36 #include "gc/g1/g1RemSet.hpp" 37 #include "gc/g1/heapRegion.inline.hpp" 38 #include "gc/g1/heapRegionManager.inline.hpp" 39 #include "gc/g1/heapRegionRemSet.hpp" 40 #include "gc/shared/gcTraceTime.inline.hpp" 41 #include "gc/shared/suspendibleThreadSet.hpp" 42 #include "memory/iterator.hpp" 43 #include "memory/resourceArea.hpp" 44 #include "oops/access.inline.hpp" 45 #include "oops/oop.inline.hpp" 46 #include "utilities/align.hpp" 47 #include "utilities/globalDefinitions.hpp" 48 #include "utilities/intHisto.hpp" 49 #include "utilities/stack.inline.hpp" 50 #include "utilities/ticks.inline.hpp" 51 52 // Collects information about the overall remembered set scan progress during an evacuation. 53 class G1RemSetScanState : public CHeapObj<mtGC> { 54 private: 55 class G1ClearCardTableTask : public AbstractGangTask { 56 G1CollectedHeap* _g1h; 57 uint* _dirty_region_list; 58 size_t _num_dirty_regions; 59 size_t _chunk_length; 60 61 size_t volatile _cur_dirty_regions; 62 public: 63 G1ClearCardTableTask(G1CollectedHeap* g1h, 64 uint* dirty_region_list, 65 size_t num_dirty_regions, 66 size_t chunk_length) : 67 AbstractGangTask("G1 Clear Card Table Task"), 68 _g1h(g1h), 69 _dirty_region_list(dirty_region_list), 70 _num_dirty_regions(num_dirty_regions), 71 _chunk_length(chunk_length), 72 _cur_dirty_regions(0) { 73 74 assert(chunk_length > 0, "must be"); 75 } 76 77 static size_t chunk_size() { return M; } 78 79 void work(uint worker_id) { 80 while (_cur_dirty_regions < _num_dirty_regions) { 81 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length; 82 size_t max = MIN2(next + _chunk_length, _num_dirty_regions); 83 84 for (size_t i = next; i < max; i++) { 85 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]); 86 if (!r->is_survivor()) { 87 r->clear_cardtable(); 88 } 89 } 90 } 91 } 92 }; 93 94 size_t _max_regions; 95 96 // Scan progress for the remembered set of a single region. Transitions from 97 // Unclaimed -> Claimed -> Complete. 98 // At each of the transitions the thread that does the transition needs to perform 99 // some special action once. This is the reason for the extra "Claimed" state. 100 typedef jint G1RemsetIterState; 101 102 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet. 103 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned. 104 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned. 105 106 G1RemsetIterState volatile* _iter_states; 107 // The current location where the next thread should continue scanning in a region's 108 // remembered set. 109 size_t volatile* _iter_claims; 110 111 // Temporary buffer holding the regions we used to store remembered set scan duplicate 112 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region) 113 uint* _dirty_region_buffer; 114 115 typedef jbyte IsDirtyRegionState; 116 static const IsDirtyRegionState Clean = 0; 117 static const IsDirtyRegionState Dirty = 1; 118 // Holds a flag for every region whether it is in the _dirty_region_buffer already 119 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools. 120 IsDirtyRegionState* _in_dirty_region_buffer; 121 size_t _cur_dirty_region; 122 123 // Creates a snapshot of the current _top values at the start of collection to 124 // filter out card marks that we do not want to scan. 125 class G1ResetScanTopClosure : public HeapRegionClosure { 126 private: 127 HeapWord** _scan_top; 128 public: 129 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { } 130 131 virtual bool do_heap_region(HeapRegion* r) { 132 uint hrm_index = r->hrm_index(); 133 if (!r->in_collection_set() && r->is_old_or_humongous()) { 134 _scan_top[hrm_index] = r->top(); 135 } else { 136 _scan_top[hrm_index] = r->bottom(); 137 } 138 return false; 139 } 140 }; 141 142 // For each region, contains the maximum top() value to be used during this garbage 143 // collection. Subsumes common checks like filtering out everything but old and 144 // humongous regions outside the collection set. 145 // This is valid because we are not interested in scanning stray remembered set 146 // entries from free or archive regions. 147 HeapWord** _scan_top; 148 public: 149 G1RemSetScanState() : 150 _max_regions(0), 151 _iter_states(NULL), 152 _iter_claims(NULL), 153 _dirty_region_buffer(NULL), 154 _in_dirty_region_buffer(NULL), 155 _cur_dirty_region(0), 156 _scan_top(NULL) { 157 } 158 159 ~G1RemSetScanState() { 160 if (_iter_states != NULL) { 161 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states); 162 } 163 if (_iter_claims != NULL) { 164 FREE_C_HEAP_ARRAY(size_t, _iter_claims); 165 } 166 if (_dirty_region_buffer != NULL) { 167 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer); 168 } 169 if (_in_dirty_region_buffer != NULL) { 170 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer); 171 } 172 if (_scan_top != NULL) { 173 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top); 174 } 175 } 176 177 void initialize(uint max_regions) { 178 assert(_iter_states == NULL, "Must not be initialized twice"); 179 assert(_iter_claims == NULL, "Must not be initialized twice"); 180 _max_regions = max_regions; 181 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC); 182 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC); 183 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC); 184 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC); 185 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC); 186 } 187 188 void reset() { 189 for (uint i = 0; i < _max_regions; i++) { 190 _iter_states[i] = Unclaimed; 191 } 192 193 G1ResetScanTopClosure cl(_scan_top); 194 G1CollectedHeap::heap()->heap_region_iterate(&cl); 195 196 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t)); 197 memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState)); 198 _cur_dirty_region = 0; 199 } 200 201 // Attempt to claim the remembered set of the region for iteration. Returns true 202 // if this call caused the transition from Unclaimed to Claimed. 203 inline bool claim_iter(uint region) { 204 assert(region < _max_regions, "Tried to access invalid region %u", region); 205 if (_iter_states[region] != Unclaimed) { 206 return false; 207 } 208 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed); 209 return (res == Unclaimed); 210 } 211 212 // Try to atomically sets the iteration state to "complete". Returns true for the 213 // thread that caused the transition. 214 inline bool set_iter_complete(uint region) { 215 if (iter_is_complete(region)) { 216 return false; 217 } 218 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed); 219 return (res == Claimed); 220 } 221 222 // Returns true if the region's iteration is complete. 223 inline bool iter_is_complete(uint region) const { 224 assert(region < _max_regions, "Tried to access invalid region %u", region); 225 return _iter_states[region] == Complete; 226 } 227 228 // The current position within the remembered set of the given region. 229 inline size_t iter_claimed(uint region) const { 230 assert(region < _max_regions, "Tried to access invalid region %u", region); 231 return _iter_claims[region]; 232 } 233 234 // Claim the next block of cards within the remembered set of the region with 235 // step size. 236 inline size_t iter_claimed_next(uint region, size_t step) { 237 return Atomic::add(step, &_iter_claims[region]) - step; 238 } 239 240 void add_dirty_region(uint region) { 241 if (_in_dirty_region_buffer[region] == Dirty) { 242 return; 243 } 244 245 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean; 246 if (marked_as_dirty) { 247 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1; 248 _dirty_region_buffer[allocated] = region; 249 } 250 } 251 252 HeapWord* scan_top(uint region_idx) const { 253 return _scan_top[region_idx]; 254 } 255 256 // Clear the card table of "dirty" regions. 257 void clear_card_table(WorkGang* workers) { 258 if (_cur_dirty_region == 0) { 259 return; 260 } 261 262 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size(); 263 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers()); 264 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion; 265 266 // Iterate over the dirty cards region list. 267 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length); 268 269 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " " 270 "units of work for " SIZE_FORMAT " regions.", 271 cl.name(), num_workers, num_chunks, _cur_dirty_region); 272 workers->run_task(&cl, num_workers); 273 274 #ifndef PRODUCT 275 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup(); 276 #endif 277 } 278 }; 279 280 G1RemSet::G1RemSet(G1CollectedHeap* g1h, 281 G1CardTable* ct, 282 G1HotCardCache* hot_card_cache) : 283 _g1h(g1h), 284 _scan_state(new G1RemSetScanState()), 285 _num_conc_refined_cards(0), 286 _ct(ct), 287 _g1p(_g1h->g1_policy()), 288 _hot_card_cache(hot_card_cache), 289 _prev_period_summary() { 290 } 291 292 G1RemSet::~G1RemSet() { 293 if (_scan_state != NULL) { 294 delete _scan_state; 295 } 296 } 297 298 uint G1RemSet::num_par_rem_sets() { 299 return DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads); 300 } 301 302 void G1RemSet::initialize(size_t capacity, uint max_regions) { 303 G1FromCardCache::initialize(num_par_rem_sets(), max_regions); 304 _scan_state->initialize(max_regions); 305 } 306 307 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state, 308 G1ScanObjsDuringScanRSClosure* scan_obj_on_card, 309 CodeBlobClosure* code_root_cl, 310 uint worker_i) : 311 _scan_state(scan_state), 312 _scan_objs_on_card_cl(scan_obj_on_card), 313 _code_root_cl(code_root_cl), 314 _strong_code_root_scan_time_sec(0.0), 315 _cards_claimed(0), 316 _cards_scanned(0), 317 _cards_skipped(0), 318 _worker_i(worker_i) { 319 _g1h = G1CollectedHeap::heap(); 320 _bot = _g1h->bot(); 321 _ct = _g1h->card_table(); 322 } 323 324 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) { 325 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card); 326 _scan_objs_on_card_cl->set_region(card_region); 327 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl); 328 _scan_objs_on_card_cl->trim_queue_partially(); 329 _cards_scanned++; 330 } 331 332 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) { 333 double scan_start = os::elapsedTime(); 334 r->strong_code_roots_do(_code_root_cl); 335 _strong_code_root_scan_time_sec += (os::elapsedTime() - scan_start); 336 } 337 338 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){ 339 _ct->set_card_claimed(card_index); 340 _scan_state->add_dirty_region(region_idx_for_card); 341 } 342 343 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) { 344 assert(r->in_collection_set(), "should only be called on elements of CS."); 345 uint region_idx = r->hrm_index(); 346 347 if (_scan_state->iter_is_complete(region_idx)) { 348 return false; 349 } 350 if (_scan_state->claim_iter(region_idx)) { 351 // If we ever free the collection set concurrently, we should also 352 // clear the card table concurrently therefore we won't need to 353 // add regions of the collection set to the dirty cards region. 354 _scan_state->add_dirty_region(region_idx); 355 } 356 357 // We claim cards in blocks so as to reduce the contention. 358 size_t const block_size = G1RSetScanBlockSize; 359 360 HeapRegionRemSetIterator iter(r->rem_set()); 361 size_t card_index; 362 363 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size); 364 for (size_t current_card = 0; iter.has_next(card_index); current_card++) { 365 if (current_card >= claimed_card_block + block_size) { 366 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size); 367 } 368 if (current_card < claimed_card_block) { 369 _cards_skipped++; 370 continue; 371 } 372 _cards_claimed++; 373 374 // If the card is dirty, then G1 will scan it during Update RS. 375 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) { 376 continue; 377 } 378 379 HeapWord* const card_start = _g1h->bot()->address_for_index(card_index); 380 uint const region_idx_for_card = _g1h->addr_to_region(card_start); 381 382 assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start), 383 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index()); 384 HeapWord* const top = _scan_state->scan_top(region_idx_for_card); 385 if (card_start >= top) { 386 continue; 387 } 388 389 // We claim lazily (so races are possible but they're benign), which reduces the 390 // number of duplicate scans (the rsets of the regions in the cset can intersect). 391 // Claim the card after checking bounds above: the remembered set may contain 392 // random cards into current survivor, and we would then have an incorrectly 393 // claimed card in survivor space. Card table clear does not reset the card table 394 // of survivor space regions. 395 claim_card(card_index, region_idx_for_card); 396 397 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top)); 398 399 scan_card(mr, region_idx_for_card); 400 } 401 if (_scan_state->set_iter_complete(region_idx)) { 402 // Scan the strong code root list attached to the current region 403 scan_strong_code_roots(r); 404 } 405 return false; 406 } 407 408 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, 409 uint worker_i) { 410 double rs_time_start = os::elapsedTime(); 411 412 CodeBlobClosure* heap_region_codeblobs = pss->closures()->weak_codeblobs(); 413 414 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss); 415 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, heap_region_codeblobs, worker_i); 416 _g1h->collection_set_iterate_from(&cl, worker_i); 417 418 double scan_rs_time_sec = (os::elapsedTime() - rs_time_start) - 419 cl.strong_code_root_scan_time_sec(); 420 421 G1GCPhaseTimes* p = _g1p->phase_times(); 422 423 double scan_rs_trim_queue_time = TicksToTimeHelper::seconds(scan_cl.trim_ticks_and_reset()); 424 425 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, scan_rs_time_sec); 426 p->move_time_secs(G1GCPhaseTimes::ScanRS, G1GCPhaseTimes::ObjCopy, worker_i, scan_rs_trim_queue_time); 427 428 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards); 429 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards); 430 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards); 431 432 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time_sec()); 433 p->move_time_secs(G1GCPhaseTimes::CodeRoots, G1GCPhaseTimes::ObjCopy, worker_i, pss->closures()->trim_time_seconds()); 434 } 435 436 // Closure used for updating rem sets. Only called during an evacuation pause. 437 class G1RefineCardClosure: public CardTableEntryClosure { 438 G1RemSet* _g1rs; 439 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl; 440 441 size_t _cards_scanned; 442 size_t _cards_skipped; 443 public: 444 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) : 445 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0) 446 {} 447 448 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 449 // The only time we care about recording cards that 450 // contain references that point into the collection set 451 // is during RSet updating within an evacuation pause. 452 // In this case worker_i should be the id of a GC worker thread. 453 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause"); 454 455 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl); 456 457 if (card_scanned) { 458 _update_rs_cl->trim_queue_partially(); 459 _cards_scanned++; 460 } else { 461 _cards_skipped++; 462 } 463 return true; 464 } 465 466 size_t cards_scanned() const { return _cards_scanned; } 467 size_t cards_skipped() const { return _cards_skipped; } 468 }; 469 470 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) { 471 G1GCPhaseTimes* p = _g1p->phase_times(); 472 double scan_hcc_trim_queue_time = 0.0; 473 474 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss, worker_i); 475 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl); 476 477 { 478 G1GCParPhaseTimesTracker x(p, G1GCPhaseTimes::UpdateRS, worker_i); 479 if (G1HotCardCache::default_use_cache()) { 480 { 481 // Apply the closure to the entries of the hot card cache. 482 G1GCParPhaseTimesTracker y(p, G1GCPhaseTimes::ScanHCC, worker_i); 483 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i); 484 } 485 scan_hcc_trim_queue_time = TicksToTimeHelper::seconds(update_rs_cl.trim_ticks_and_reset()); 486 p->move_time_secs(G1GCPhaseTimes::ScanHCC, G1GCPhaseTimes::ObjCopy, worker_i, scan_hcc_trim_queue_time); 487 } 488 // Apply the closure to all remaining log entries. 489 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i); 490 } 491 double update_rs_trim_queue_time = TicksToTimeHelper::seconds(update_rs_cl.trim_ticks_and_reset()); 492 493 p->move_time_secs(G1GCPhaseTimes::UpdateRS, G1GCPhaseTimes::ObjCopy, worker_i, update_rs_trim_queue_time + scan_hcc_trim_queue_time); 494 495 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards); 496 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards); 497 } 498 499 void G1RemSet::cleanupHRRS() { 500 HeapRegionRemSet::cleanup(); 501 } 502 503 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) { 504 update_rem_set(pss, worker_i); 505 scan_rem_set(pss, worker_i);; 506 } 507 508 void G1RemSet::prepare_for_oops_into_collection_set_do() { 509 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 510 dcqs.concatenate_logs(); 511 512 _scan_state->reset(); 513 } 514 515 void G1RemSet::cleanup_after_oops_into_collection_set_do() { 516 G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times(); 517 518 // Set all cards back to clean. 519 double start = os::elapsedTime(); 520 _scan_state->clear_card_table(_g1h->workers()); 521 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0); 522 } 523 524 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) { 525 #ifdef ASSERT 526 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 527 assert(g1h->is_in_exact(ct->addr_for(card_ptr)), 528 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap", 529 p2i(card_ptr), 530 ct->index_for(ct->addr_for(card_ptr)), 531 p2i(ct->addr_for(card_ptr)), 532 g1h->addr_to_region(ct->addr_for(card_ptr))); 533 #endif 534 } 535 536 void G1RemSet::refine_card_concurrently(jbyte* card_ptr, 537 uint worker_i) { 538 assert(!_g1h->is_gc_active(), "Only call concurrently"); 539 540 check_card_ptr(card_ptr, _ct); 541 542 // If the card is no longer dirty, nothing to do. 543 if (*card_ptr != G1CardTable::dirty_card_val()) { 544 return; 545 } 546 547 // Construct the region representing the card. 548 HeapWord* start = _ct->addr_for(card_ptr); 549 // And find the region containing it. 550 HeapRegion* r = _g1h->heap_region_containing(start); 551 552 // This check is needed for some uncommon cases where we should 553 // ignore the card. 554 // 555 // The region could be young. Cards for young regions are 556 // distinctly marked (set to g1_young_gen), so the post-barrier will 557 // filter them out. However, that marking is performed 558 // concurrently. A write to a young object could occur before the 559 // card has been marked young, slipping past the filter. 560 // 561 // The card could be stale, because the region has been freed since 562 // the card was recorded. In this case the region type could be 563 // anything. If (still) free or (reallocated) young, just ignore 564 // it. If (reallocated) old or humongous, the later card trimming 565 // and additional checks in iteration may detect staleness. At 566 // worst, we end up processing a stale card unnecessarily. 567 // 568 // In the normal (non-stale) case, the synchronization between the 569 // enqueueing of the card and processing it here will have ensured 570 // we see the up-to-date region type here. 571 if (!r->is_old_or_humongous()) { 572 return; 573 } 574 575 // The result from the hot card cache insert call is either: 576 // * pointer to the current card 577 // (implying that the current card is not 'hot'), 578 // * null 579 // (meaning we had inserted the card ptr into the "hot" card cache, 580 // which had some headroom), 581 // * a pointer to a "hot" card that was evicted from the "hot" cache. 582 // 583 584 if (_hot_card_cache->use_cache()) { 585 assert(!SafepointSynchronize::is_at_safepoint(), "sanity"); 586 587 const jbyte* orig_card_ptr = card_ptr; 588 card_ptr = _hot_card_cache->insert(card_ptr); 589 if (card_ptr == NULL) { 590 // There was no eviction. Nothing to do. 591 return; 592 } else if (card_ptr != orig_card_ptr) { 593 // Original card was inserted and an old card was evicted. 594 start = _ct->addr_for(card_ptr); 595 r = _g1h->heap_region_containing(start); 596 597 // Check whether the region formerly in the cache should be 598 // ignored, as discussed earlier for the original card. The 599 // region could have been freed while in the cache. 600 if (!r->is_old_or_humongous()) { 601 return; 602 } 603 } // Else we still have the original card. 604 } 605 606 // Trim the region designated by the card to what's been allocated 607 // in the region. The card could be stale, or the card could cover 608 // (part of) an object at the end of the allocated space and extend 609 // beyond the end of allocation. 610 611 // Non-humongous objects are only allocated in the old-gen during 612 // GC, so if region is old then top is stable. Humongous object 613 // allocation sets top last; if top has not yet been set, this is 614 // a stale card and we'll end up with an empty intersection. If 615 // this is not a stale card, the synchronization between the 616 // enqueuing of the card and processing it here will have ensured 617 // we see the up-to-date top here. 618 HeapWord* scan_limit = r->top(); 619 620 if (scan_limit <= start) { 621 // If the trimmed region is empty, the card must be stale. 622 return; 623 } 624 625 // Okay to clean and process the card now. There are still some 626 // stale card cases that may be detected by iteration and dealt with 627 // as iteration failure. 628 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val(); 629 630 // This fence serves two purposes. First, the card must be cleaned 631 // before processing the contents. Second, we can't proceed with 632 // processing until after the read of top, for synchronization with 633 // possibly concurrent humongous object allocation. It's okay that 634 // reading top and reading type were racy wrto each other. We need 635 // both set, in any order, to proceed. 636 OrderAccess::fence(); 637 638 // Don't use addr_for(card_ptr + 1) which can ask for 639 // a card beyond the heap. 640 HeapWord* end = start + G1CardTable::card_size_in_words; 641 MemRegion dirty_region(start, MIN2(scan_limit, end)); 642 assert(!dirty_region.is_empty(), "sanity"); 643 644 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i); 645 646 bool card_processed = 647 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl); 648 649 // If unable to process the card then we encountered an unparsable 650 // part of the heap (e.g. a partially allocated object) while 651 // processing a stale card. Despite the card being stale, redirty 652 // and re-enqueue, because we've already cleaned the card. Without 653 // this we could incorrectly discard a non-stale card. 654 if (!card_processed) { 655 // The card might have gotten re-dirtied and re-enqueued while we 656 // worked. (In fact, it's pretty likely.) 657 if (*card_ptr != G1CardTable::dirty_card_val()) { 658 *card_ptr = G1CardTable::dirty_card_val(); 659 MutexLockerEx x(Shared_DirtyCardQ_lock, 660 Mutex::_no_safepoint_check_flag); 661 DirtyCardQueue* sdcq = 662 JavaThread::dirty_card_queue_set().shared_dirty_card_queue(); 663 sdcq->enqueue(card_ptr); 664 } 665 } else { 666 _num_conc_refined_cards++; // Unsynchronized update, only used for logging. 667 } 668 } 669 670 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr, 671 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) { 672 assert(_g1h->is_gc_active(), "Only call during GC"); 673 674 check_card_ptr(card_ptr, _ct); 675 676 // If the card is no longer dirty, nothing to do. This covers cards that were already 677 // scanned as parts of the remembered sets. 678 if (*card_ptr != G1CardTable::dirty_card_val()) { 679 return false; 680 } 681 682 // We claim lazily (so races are possible but they're benign), which reduces the 683 // number of potential duplicate scans (multiple threads may enqueue the same card twice). 684 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val(); 685 686 // Construct the region representing the card. 687 HeapWord* card_start = _ct->addr_for(card_ptr); 688 // And find the region containing it. 689 uint const card_region_idx = _g1h->addr_to_region(card_start); 690 691 _scan_state->add_dirty_region(card_region_idx); 692 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx); 693 if (scan_limit <= card_start) { 694 // If the card starts above the area in the region containing objects to scan, skip it. 695 return false; 696 } 697 698 // Don't use addr_for(card_ptr + 1) which can ask for 699 // a card beyond the heap. 700 HeapWord* card_end = card_start + G1CardTable::card_size_in_words; 701 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end)); 702 assert(!dirty_region.is_empty(), "sanity"); 703 704 HeapRegion* const card_region = _g1h->region_at(card_region_idx); 705 update_rs_cl->set_region(card_region); 706 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl); 707 assert(card_processed, "must be"); 708 return true; 709 } 710 711 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) { 712 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) && 713 (period_count % G1SummarizeRSetStatsPeriod == 0)) { 714 715 G1RemSetSummary current(this); 716 _prev_period_summary.subtract_from(¤t); 717 718 Log(gc, remset) log; 719 log.trace("%s", header); 720 ResourceMark rm; 721 LogStream ls(log.trace()); 722 _prev_period_summary.print_on(&ls); 723 724 _prev_period_summary.set(¤t); 725 } 726 } 727 728 void G1RemSet::print_summary_info() { 729 Log(gc, remset, exit) log; 730 if (log.is_trace()) { 731 log.trace(" Cumulative RS summary"); 732 G1RemSetSummary current(this); 733 ResourceMark rm; 734 LogStream ls(log.trace()); 735 current.print_on(&ls); 736 } 737 } 738 739 class G1RebuildRemSetTask: public AbstractGangTask { 740 // Aggregate the counting data that was constructed concurrently 741 // with marking. 742 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure { 743 G1ConcurrentMark* _cm; 744 G1RebuildRemSetClosure _update_cl; 745 746 // Applies _update_cl to the references of the given object, limiting objArrays 747 // to the given MemRegion. Returns the amount of words actually scanned. 748 size_t scan_for_references(oop const obj, MemRegion mr) { 749 size_t const obj_size = obj->size(); 750 // All non-objArrays and objArrays completely within the mr 751 // can be scanned without passing the mr. 752 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) { 753 obj->oop_iterate(&_update_cl); 754 return obj_size; 755 } 756 // This path is for objArrays crossing the given MemRegion. Only scan the 757 // area within the MemRegion. 758 obj->oop_iterate(&_update_cl, mr); 759 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size(); 760 } 761 762 // A humongous object is live (with respect to the scanning) either 763 // a) it is marked on the bitmap as such 764 // b) its TARS is larger than TAMS, i.e. has been allocated during marking. 765 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const { 766 return bitmap->is_marked(humongous_obj) || (tars > tams); 767 } 768 769 // Iterator over the live objects within the given MemRegion. 770 class LiveObjIterator : public StackObj { 771 const G1CMBitMap* const _bitmap; 772 const HeapWord* _tams; 773 const MemRegion _mr; 774 HeapWord* _current; 775 776 bool is_below_tams() const { 777 return _current < _tams; 778 } 779 780 bool is_live(HeapWord* obj) const { 781 return !is_below_tams() || _bitmap->is_marked(obj); 782 } 783 784 HeapWord* bitmap_limit() const { 785 return MIN2(const_cast<HeapWord*>(_tams), _mr.end()); 786 } 787 788 void move_if_below_tams() { 789 if (is_below_tams() && has_next()) { 790 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 791 } 792 } 793 public: 794 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) : 795 _bitmap(bitmap), 796 _tams(tams), 797 _mr(mr), 798 _current(first_oop_into_mr) { 799 800 assert(_current <= _mr.start(), 801 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")", 802 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end())); 803 804 // Step to the next live object within the MemRegion if needed. 805 if (is_live(_current)) { 806 // Non-objArrays were scanned by the previous part of that region. 807 if (_current < mr.start() && !oop(_current)->is_objArray()) { 808 _current += oop(_current)->size(); 809 // We might have positioned _current on a non-live object. Reposition to the next 810 // live one if needed. 811 move_if_below_tams(); 812 } 813 } else { 814 // The object at _current can only be dead if below TAMS, so we can use the bitmap. 815 // immediately. 816 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 817 assert(_current == _mr.end() || is_live(_current), 818 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")", 819 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end())); 820 } 821 } 822 823 void move_to_next() { 824 _current += next()->size(); 825 move_if_below_tams(); 826 } 827 828 oop next() const { 829 oop result = oop(_current); 830 assert(is_live(_current), 831 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d", 832 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result)); 833 return result; 834 } 835 836 bool has_next() const { 837 return _current < _mr.end(); 838 } 839 }; 840 841 // Rebuild remembered sets in the part of the region specified by mr and hr. 842 // Objects between the bottom of the region and the TAMS are checked for liveness 843 // using the given bitmap. Objects between TAMS and TARS are assumed to be live. 844 // Returns the number of live words between bottom and TAMS. 845 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap, 846 HeapWord* const top_at_mark_start, 847 HeapWord* const top_at_rebuild_start, 848 HeapRegion* hr, 849 MemRegion mr) { 850 size_t marked_words = 0; 851 852 if (hr->is_humongous()) { 853 oop const humongous_obj = oop(hr->humongous_start_region()->bottom()); 854 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) { 855 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start); 856 // however in case of humongous objects it is sufficient to scan the encompassing 857 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the 858 // two areas will be zero sized. I.e. TAMS is either 859 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different 860 // value: this would mean that TAMS points somewhere into the object. 861 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start, 862 "More than one object in the humongous region?"); 863 humongous_obj->oop_iterate(&_update_cl, mr); 864 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0; 865 } else { 866 return 0; 867 } 868 } 869 870 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) { 871 oop obj = it.next(); 872 size_t scanned_size = scan_for_references(obj, mr); 873 if ((HeapWord*)obj < top_at_mark_start) { 874 marked_words += scanned_size; 875 } 876 } 877 878 return marked_words * HeapWordSize; 879 } 880 public: 881 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h, 882 G1ConcurrentMark* cm, 883 uint worker_id) : 884 HeapRegionClosure(), 885 _cm(cm), 886 _update_cl(g1h, worker_id) { } 887 888 bool do_heap_region(HeapRegion* hr) { 889 if (_cm->has_aborted()) { 890 return true; 891 } 892 893 uint const region_idx = hr->hrm_index(); 894 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);) 895 assert(top_at_rebuild_start_check == NULL || 896 top_at_rebuild_start_check > hr->bottom(), 897 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)", 898 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str()); 899 900 size_t total_marked_bytes = 0; 901 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize; 902 903 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start(); 904 905 HeapWord* cur = hr->bottom(); 906 while (cur < hr->end()) { 907 // After every iteration (yield point) we need to check whether the region's 908 // TARS changed due to e.g. eager reclaim. 909 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx); 910 if (top_at_rebuild_start == NULL) { 911 return false; 912 } 913 914 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words)); 915 if (next_chunk.is_empty()) { 916 break; 917 } 918 919 const Ticks start = Ticks::now(); 920 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(), 921 top_at_mark_start, 922 top_at_rebuild_start, 923 hr, 924 next_chunk); 925 Tickspan time = Ticks::now() - start; 926 927 log_trace(gc, remset, tracking)("Rebuilt region %u " 928 "live " SIZE_FORMAT " " 929 "time %.3fms " 930 "marked bytes " SIZE_FORMAT " " 931 "bot " PTR_FORMAT " " 932 "TAMS " PTR_FORMAT " " 933 "TARS " PTR_FORMAT, 934 region_idx, 935 _cm->liveness(region_idx) * HeapWordSize, 936 TicksToTimeHelper::seconds(time) * 1000.0, 937 marked_bytes, 938 p2i(hr->bottom()), 939 p2i(top_at_mark_start), 940 p2i(top_at_rebuild_start)); 941 942 if (marked_bytes > 0) { 943 total_marked_bytes += marked_bytes; 944 } 945 cur += chunk_size_in_words; 946 947 _cm->do_yield_check(); 948 if (_cm->has_aborted()) { 949 return true; 950 } 951 } 952 // In the final iteration of the loop the region might have been eagerly reclaimed. 953 // Simply filter out those regions. We can not just use region type because there 954 // might have already been new allocations into these regions. 955 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);) 956 assert(top_at_rebuild_start == NULL || 957 total_marked_bytes == hr->marked_bytes(), 958 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " " 959 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")", 960 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(), 961 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start)); 962 // Abort state may have changed after the yield check. 963 return _cm->has_aborted(); 964 } 965 }; 966 967 HeapRegionClaimer _hr_claimer; 968 G1ConcurrentMark* _cm; 969 970 uint _worker_id_offset; 971 public: 972 G1RebuildRemSetTask(G1ConcurrentMark* cm, 973 uint n_workers, 974 uint worker_id_offset) : 975 AbstractGangTask("G1 Rebuild Remembered Set"), 976 _cm(cm), 977 _hr_claimer(n_workers), 978 _worker_id_offset(worker_id_offset) { 979 } 980 981 void work(uint worker_id) { 982 SuspendibleThreadSetJoiner sts_join; 983 984 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 985 986 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id); 987 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id); 988 } 989 }; 990 991 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm, 992 WorkGang* workers, 993 uint worker_id_offset) { 994 uint num_workers = workers->active_workers(); 995 996 G1RebuildRemSetTask cl(cm, 997 num_workers, 998 worker_id_offset); 999 workers->run_task(&cl, num_workers); 1000 }